Fat-Tree Routing for Transit
Abstract
As an alternative to using a bidelta network topology for large Transit
networks, I consider the requirements to extend the base Transit network
into Leiserson's Fat-Tree configuration. Transit will be a high-speed,
low-latency, fault-tolerant network interconnection for high performance
multi-processor computers. The initial interconnect scheme planned for
Transit will use a bidelta style network to support up to 256 processors.
Scaling beyond 256 processors by simply extending that network topology
will result in a uniform degradation of network latency across processors.
A fat-tree network structure will allow the Transit network to be scaled
arbitrarily while taking advantage of the locality and universality of
fat-trees to minimize the impact of scaling upon network latency. I
consider the topology and construction issues for integrating the Transit
routing network component and technology into a fat-tree configuration. I
also characterize the resulting network's size, locality, and performance
and compare these characteristics with those of bidelta networks.
André DeHon <andre@mit.edu>
MIT
Transit Project
MIT
AI Lab